Multimodal AI Boosts Zero-Shot Drug Discovery

Drug discovery often relies on multimodal data, combining chemical structures with cellular responses like gene expression and cell morphology. However, directly fusing these diverse data types can introduce noise or incorrectly separate biologically related compounds, hindering the discovery of transferable mechanism patterns for new drugs. A new framework, PMRD (Pharmacological Response Domain-guided), addresses these challenges by separating mechanism-consistent factors from modality-specific noise. It constructs a "consensus response domain" across three different modalities, ensuring that the learned drug representations are more biologically coherent. PMRD employs mechanism candidate augmentation to identify stable factors and uses retrieval-geometry attribution to dynamically adjust alignment and augmentation objectives. This feedback mechanism suppresses training signals that conflict with mechanism-discriminative retrieval. By combining complementary representations through reliability-aware multiview retrieval, PMRD significantly improves zero-shot property prediction on public datasets and creates more biologically meaningful drug neighborhoods, even for structurally dissimilar but functionally related compounds.